Research topic: Active learning of dynamic systems with safety bounds
The submitted paper is entitled "Efficiently Computable Safety Bounds for Gaussian Processes in Active Learning" and deals with the active learning of dynamic systems where safety constraints must be met. In technical applications, the design space is often explored using continuous trajectories along which safety must be assessed. Gaussian processes (GP) are used for uncertainty estimation. Especially for strict safety requirements, GP methods require computationally intensive Monte Carlo sampling procedures.
The presented method: Provable safety bounds by adaptively sampled median of the posterior GP
The paper presents a method that provides provable estimates of safety based on the adaptively simulated median of the a-posteriori GP. This method significantly reduces the number of simulated samples required, leading to much faster evaluations of the safety of potential trajectories without compromising accuracy and exploration speed. Overall, this promises faster exploration of the design space, saving valuable resources such as time and labour costs. The effectiveness of the approach will be demonstrated through simulations and validation on a real engine.
The collaboration of all individuals and institutions for the jointly published paper emphasises the innovative power of interdisciplinary cooperation between industry and research.
About AISTATS
AISTATS is a prestigious event with over 500 accepted papers and almost 1,000 participants from all over the world. AISTATS is an interdisciplinary association of researchers at the intersection of computer science, artificial intelligence, machine learning, statistics and related fields.
You can read the paper here:https://proceedings.mlr.press/v238/tebbe24a.html